Keywords

IPP, MIPP, Reinforcement Learning, DQN, Machine Learning

Abstract

Informative path planning (IPP) algorithms are widely used to control the movement of drones or ground robots when the objective is to efficiently collect information from an environment of interest. For instance, in agriculture, drones might be used for the timely detection of outbreaks of plant diseases. Two distinct classes of IPP algorithms are the system2atic algorithms, where the robot moves in a regular planned pattern (such as lawnmower) aiming for uniform sampling and random algorithms (such as random waypoint) that aims to collect a random sample of the environment. In this dissertation, we present several novel IPP algorithms, that specifically adapt to the needs of precision agriculture, where the value of information for certain type of observations (detecting disease) is much higher than other types ({\em e.g.} confirmation of healthy plants). The first contribution is finding a hybrid offline IPP algorithm that combines the strengths of the systematic and the random algorithms. The second work proposes an innovative adaptive IPP algorithm that incorporates reinforcement learning (RL) algorithm that aims to move the robot to locations that increase environmental information gain based on the robot's perception of the environment. While the third contribution employs an approximate RL algorithm using Deep Q-Learning Network (DQN) that supports a more complex state representation needed to achieve better IPP performance. The final work proposes a solution to the multi-robot informative path planning (MIPP) by using a DQN model for close inspection and a learning-based dispatch model to deploy robots in a way that increases the likelihood of anomaly discovery in the environment, given a time constraint.

Completion Date

2025

Semester

Fall

Committee Chair

Ladislau Boloni

Degree

Doctor of Philosophy (Ph.D.)

College

College of Engineering and Computer Science

Department

Computer Science

Format

PDF

Identifier

DP0029802

Document Type

Thesis

Campus Location

Orlando (Main) Campus

Download

Share

COinS